KR102080362B1 - Query expansion - Google Patents

Abstract

The present disclosure provides an example query extension method and system. The query entered by the user is received. The normalized query of a query is determined according to that query. Normalized queries are used as an extension term in the query to enforce query expansion. For example, session information is obtained from a user's search log. All queries appearing in a single session are obtained and the vote count is counted or counted for each query. Vote similarity between the single query and the target query is determined, and the degree of correlation between the single query and the target query is determined according to the vote similarity. The normalized query of the target query is determined by the degree of correlation. This technique accurately scales queries and reduces query time, improving system response speed and processing efficiency.

Description

Query extension {QUERY EXPANSION}

Cross Reference of Related Patent Application

This application claims the foreign priority of Chinese Patent Application No. 201210254810.0, filed July 20, 2012, entitled "Query Expansion Method and System," which is incorporated by reference in its entirety. .

Technical Field

FIELD The present disclosure relates to the field of computer data processing, and more particularly, to query expansion methods and systems.

With the development of network technology, search engines have been continuously improved and various information can be obtained from the Internet through these search engines. Search engines have become one of the main ways to help users quickly find information on the Internet. The user submits a query term (query) to the search engine, which returns a search result corresponding to the query to the user.

In an e-commerce website, more particularly in a large e-commerce website, a user usually has to search and find his / her desired product using a query. Since the queries entered by the user are usually constructed according to the user's wishes, the results corresponding to the query may be high or low, resulting in low accuracy and frequent searches of the search results. Thus, search engines can usually expand or rewrite queries to enrich query information and intelligently optimize queries entered by users, thereby improving the accuracy of search results and at the same time frequenting searches by users. This reduces the pressure on the server.

Typical query extension methods include extensions at a query end and at an index end. Extensions at the query end mainly include adding, replacing, and deleting queries. That is, certain characters or portions are added, replaced or deleted from the query entered by the user. For example, if the query entered by the user is a "Nokia ™ mobile phone", the add action may apply to the query to change the query to "Nokia ™ N95 mobile phone", and the delete action may change the query to "Nokia". ™ "or" mobile phone "can be applied to the query, or alternative actions can be applied to the query to change the query to" Samsung ™ mobile phone "or" Apple ™ mobile phone "and the like. Expansion at the index end refers primarily to synonym expansion of queries at the index end. The set of synonyms is usually obtained through normal data mining. When a term appears, its synonyms are extracted from the set of synonyms for expansion. In order to ensure the accuracy of the search results and the query, concurrent extensions may be employed at the query end and the index end. In other words, the query is expanded on both the query end and the index end, respectively, and the result corresponding to the same extension term is selected as the extension result.

During actual processing, the search engine usually selects extended terms one by one at the query end to search according to a particular sequence, matches those extended terms with extended terms at the index end, and if there is a match between the extended terms, the extended terms Can return search results. During this process, it is possible for a query end to have multiple extended terms, but only one extended term at the index end. So, depending on the sequence, the last extension term at the query end matches the query term at the index end. The search engine must search multiple times until the last extended term at the query end matches the extended term at the index end. Thus, not only the invalid search time of the search engine but also the time for the system to return the search results, the system response speed is reduced, and the system resource share is increased.

This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify all key features or essential features of the claimed subject matter, nor is it intended to be used only as an aid in determining the scope of the claimed subject matter. For example, the term "technology" may refer to device (s), system (s), method (s) and / or computer-readable instructions as permitted by the context of the foregoing and entirety of this disclosure.

The present disclosure provides a query extension method and system. For example, the present technology can solve the problem of excessive search time and system resource occupancy, which also affects system response speed and efficiency, caused by repeated search and matching of queries by search engines.

This disclosure describes an example query extension method. The query entered by the user is received. The normalized query of a query is determined by the query. Normalized queries are used as an extension term in the query to enforce query expansion.

The normalized query of the query can be determined using the following method. Session information in the user's search log is obtained. All queries appearing in a single session are obtained and the vote count is counted or counted for each query. In a single session, depending on the appearance sequence of each query, any query that appears before a particular query is counted as one vote (for that particular query).

The degree of vote similarity between a single query and a target query is determined by the total number of votes in the target query and the number of votes in a single query for the target query in all sessions. The degree of correlation between a single query and the target query is determined by the degree of vote similarity. The normalized query of the target query is determined by the degree of correlation between the single query and the target query.

For example, the total number of votes for the target query in all sessions is calculated or counted as follows. One or more sessions containing the target query are obtained. The vote count of the target query is counted in each session. In each session, the number of votes is accumulated to obtain the total number of votes for the target query.

For example, the number of votes in a single query for a target query is calculated or counted as follows. One or more sessions are obtained that include a single query and a target query. It is determined whether a single query provides a vote for the target query in each session. If the decision is positive, then each session is selected. The number of selected sessions is calculated to obtain the number of votes of a single query for the target query.

For example, the degree of vote similarity between a single query and the target query may be determined as follows. The ratio of the total number of votes in a single query to the target query to the total vote scores of all the queries for the target query (for a single product) is used as the degree of vote similarity between the single query and the target query.

As another example, the degree of vote similarity between a single query and a target query may also be determined as follows. The weight and a base number of each vote for the target query is determined. The score of each vote is calculated according to the weight and the cardinality. The ratio of the total vote score of a single query to the target query to the total vote score of all the queries for the target query is used as the degree of vote similarity between the single query and the target query.

For example, the normalized query of the target query may be determined as follows according to the degree of correlation between the single query and the target query. The threshold of the normalized query is set. If the value of the degree of correlation between the single query and the target query exceeds the threshold of the normalized query, the single query is determined as the normalized query of the target query.

As another example, the normalized query of the target query may also be determined as follows according to the degree of correlation between the single query and the target query. The categories of normalized queries are separated into synonym normalized queries, correlated normalized queries, and extended normalized queries. The range of values for the three categories is set in descending order, respectively, according to the degree of correlation. The category corresponding to the value range to which the degree of correlation between the single query and the target query belongs is used as the detailed category of the single query and the target query.

For example, before the degree of correlation between a single query and the target query is determined according to the degree of vote similarity, the method may further include the following operation. Click information of the search results in the user's search log is obtained. Search results including the target query are extracted from the click information. The degree of click similarity between the single query and the target query is determined according to the total number of clicks of the search results including the target query and the total clicks of the search results corresponding to the single query while including the target query.

For example, the degree of correlation between a single query and the target query is determined according to the degree of vote similarity and the degree of click similarity.

For example, the degree of correlation between the single query and the target query may be determined as follows according to the degree of vote similarity and the degree of click similarity. The larger value between the degree of vote similarity and the degree of click similarity is used as the degree of correlation between the target query and the target query.

Alternatively, weights of vote similarity and click similarity are determined. The degree of correlation between the single query and the target query is calculated according to one or more predetermined rules based on the degree of vote similarity and the degree of click similarity as well as their respective weights.

For example, prior to the step of determining the degree of correlation between a single query and the target query according to the degree of vote similarity, the method further includes the following operation. Merchant data stored in the server is obtained. Seller data includes product description information specified when the seller describes the product. Seller data is analyzed and query and characteristic terms of the query are extracted from the seller data. The degree of feature similarity is determined by the feature terminology of the single query and the target query.

For example, the degree of correlation between a single query and a target query can be determined as follows. The degree of correlation between a single query and the target query may be determined according to the degree of vote similarity and the degree of feature similarity.

For example, the degree of correlation between the single query and the target query may be determined as follows according to the degree of vote similarity and the degree of feature similarity. The characteristic value of each characteristic term is calculated. The characteristic value is calculated according to the click information of the characteristic term and its corresponding query. The degree of feature similarity between a single query and the target query is calculated according to the property value.

For example, prior to the step of determining a normalized query of the target query according to the degree of correlation between the single query and the target query, the method may further include the following operation. The degree of semantic similarity and / or category similarity between a single query and the target query is determined.

The normalized query of the target query can be determined as follows according to the degree of correlation between the single query and the target query. The normalized query of the target query may be determined according to the degree of correlation and semantic similarity between the single query and the target query. Alternatively, the normalized query of the target query can be determined according to the degree of correlation and category similarity between the single query and the target query. Alternatively, the normalized query of the target query may be determined according to the degree of correlation, semantic similarity, and category similarity between the single query and the target query.

For example, the determination of semantic similarity between a single query and a target query may include the following actions. The editing distance between the single query and the target query is determined. The editing distance refers to the minimum number of editing operations for converting from one term to another term. The edit distance is normalized to obtain a degree of semantic similarity with a quantitative level equal to the degree of correlation.

The present disclosure also provides an example query extension system. The system can include a query input module, a normalized query decision module, and a query extension module.

The query input module obtains a query entered by the user. The normalized query decision module determines the normalized query of the query according to the query. The query expansion module implements query expansion using normalized queries as the query's extension term.

The normalized query determination module may include a session information acquisition module, a query vote calculation module, a vote similarity degree determination module, a correlation degree determination module, and a normalized query determination module.

The session information obtaining module obtains session information from the search log of the user.

The query vote calculation module obtains all queries appearing in a single session and counts votes for each query. In a single session, depending on the appearance sequence of each query, any query that appears before a particular query is counted as one vote (for that particular query).

The vote similarity determination module determines the degree of vote similarity between the single query and the target query according to the total number of votes of the target query and the number of votes of the single query for the target query in all sessions.

The degree of correlation determination module determines the degree of correlation between a single query and the target query according to the degree of vote similarity.

The normalized query determination module determines the normalized query of the target query based on the degree of correlation between the single query and the target query.

For example, the vote similarity degree determination module may include a cardinal and weight determination unit, a score calculation unit, and a ratio calculation unit. The cardinal and weight determination unit determines the weight and cardinality of each vote for the target query. The score calculation unit calculates the score of each vote according to the weight and the cardinality. The ratio calculation unit uses the ratio of the total vote score of a single query to the target query to the total vote score of all the queries for the target query as the degree of vote similarity between the single query and the target query.

For example, the normalized query decision module sets the threshold of a normalized query and determines that a single query as the normalized query of the target query if the correlation degree value between the single query and the target query exceeds the threshold of the normalized query. It may include a normalized query threshold setting unit.

As another example, the normalized query determination module may include a normalized query category classification unit, a value range setting unit, and a category determination unit. The normalized query category classification unit separates the normalized query categories into synonym normalized queries, correlated normalized queries, and extended normalized queries.

The value range setting unit sets the value ranges of the three categories in descending order according to the correlation degree value.

The category determining unit determines the category corresponding to the range of values to which the degree of correlation between the single query and the target query belongs as a detailed category of the single query and the target query.

For example, the normalized query determination module may include a click information acquisition module, a search result extraction module, and a click similarity degree determination module. The click information obtaining module obtains click information of a search result from a user's search log. The search result extraction module extracts a search result including the target query from the click information. The click similarity determining module determines the degree of click similarity between a single query and the target query according to the total number of clicks of the search result including the target query and the total number of clicks of the search result corresponding to the single query while including the target query. The degree of correlation determination module determines the degree of correlation between the single query and the target query according to the degree of vote similarity and the degree of click similarity.

For example, the normalized query determination module may also include a merchant data acquisition module, a data analysis module, and a characteristic similarity degree determination module.

The seller data acquisition module acquires seller data stored in the server. Seller data includes product description information specified when the seller describes the product.

The data analysis module analyzes the seller data and extracts query terminology as well as queries from the seller data. The degree of characteristic similarity determination module determines characteristic similarity according to characteristic terms of a single query and a target query. The degree of correlation determination module determines the degree of correlation between a single query and a target query according to the degree of vote similarity and the degree of similarity of characteristics.

For example, the characteristic similarity determining module may include a characteristic value calculation unit that calculates characteristic similarity between a single query and a target query according to the characteristic value and the characteristic value of each characteristic term. The characteristic value is calculated according to the click information of the characteristic term and its corresponding query.

As another example, the normalized query determination module may also include a semantic similarity degree determination module and / or a category similarity degree determination module that respectively determine a degree of semantic similarity and / or category similarity between a single query and a target query.

The normalized query determination module determines the normalized query of the target query according to the degree of correlation and semantic similarity between the single query and the target query. Alternatively, the normalized query determination module determines the normalized query of the target query according to the degree of correlation and category similarity between the single query and the target query. Alternatively, the normalized query determination module determines the normalized query of the target query according to the degree of correlation, degree of semantic similarity, and category similarity between the single query and the target query.

For example, the semantic similarity degree determination module may include an edit distance determination unit and a normalization processing unit. The edit distance determining unit determines the edit distance between a single query and the target query. The editing distance refers to the minimum number of editing operations for converting one term into another term. The normalization processing unit normalizes the edit distance to obtain a degree of semantic similarity having a quantitative level equal to the degree of correlation.

The query extension method and system of the present disclosure normalizes a query by utilizing session information in a search log of a user, and treats a query having the same or similar meaning as the same or similar query. After acquiring the query, the technology can automatically normalize the query and simultaneously expand the search, and ensure the accuracy of the search results at the same time as it increases the scope of the search results. The technique performs normalization by using session information in a user's search log, extracts a query included in each session, and analyzes user behavior according to a sequence of queries, thereby obtaining a process of changing a user's query during a search. Can be. Since a single session records user search information in a shorter, continuous time, the degree of correlation between queries in a single session can be greater. Therefore, extended query processing is handled based on such characteristics of session information to improve the effectiveness of normalization processing and to ensure a sufficiently high correlation between the two queries after normalization, ensuring the accuracy of the final search results and reducing the search time. Can be. Thus, system resource occupancy is reduced and system response speed and query expansion efficiency are improved.

In addition, based on the nature of the session information, the probability that queries with a small degree of correlation appear in the same session is reduced. Thus, the data range to be normalized is reduced. Processing speed is improved and processing time is saved.

Furthermore, in addition to considering session information for processing, the present technology can further improve the accuracy of normalization processing by further considering other dimensions such as user click information and seller data.

Obviously, any product of the present disclosure need not have all of the foregoing features.

To better describe the embodiments of the present disclosure, the drawings to be used in the description of the embodiments are briefly introduced as follows. Obviously, the following figures are only related to some embodiments of the present disclosure. Those skilled in the art may obtain other drawings according to the drawings in the present disclosure without creative efforts.
1 illustrates a flowchart of an example query extension method according to the first exemplary embodiment of the present disclosure.
2 illustrates a flowchart of an example method of determining a normalized query of a query according to the first example embodiment of the present disclosure.
3 shows a flowchart of another example method of determining a normalized query of a query according to the second example embodiment of the present disclosure.
4 shows a flowchart of another example method of determining a normalized query of a query according to the third example embodiment of the present disclosure.
5 shows a diagram of an example query extension system in accordance with the present disclosure.
6 shows a diagram of a first exemplary normalized query decision module in accordance with the present disclosure.
7 shows a diagram of a second exemplary normalized query determination module in accordance with this disclosure.
8 shows a diagram of a third exemplary normalized query decision module in accordance with the present disclosure.

BRIEF DESCRIPTION OF THE DRAWINGS In order to make the object, features and advantages of the present disclosure clear and easy to understand, the following description is described with reference to the drawings and some exemplary embodiments.

In this disclosure, a query may be a key term entered by a user to query and obtain expected results. For example, the query may include a product name, product brand, product model, or other term. In a particular field, a query may be a term of a particular category. For example, when using an e-commerce website, a query may be a product term that represents a product name or category, such as a mobile phone, a dress, or the like. Product terms are public queries because product terms improve the relative and effective degree of matching between search results and user expectations.

For example, by analyzing a user's search log, about 57 percent of queries are retrieved using product terms and nearly 88 percent of queries contain product terms. In addition, on an e-commerce website, a product supplier can describe a product and store the description on a server. The description may include a product name and a detailed description of the product. In a general search method, a search engine of an e-commerce website matches a product term entered by a user with a product name in a server to obtain a search result according to the matching result. Therefore, correlating product terms entered by the user with product names stored on the server by the product provider is an important premise to improve the accuracy of the search results.

Also, the amount of information and data on some large websites is huge. However, the queries included in the informational data are much less than the total informational data. Thus, if a query is normalized and a query expressing the same or similar semantics is correlated and considered as the same or similar during the query, data redundancy can be further reduced and the response speed of the search engine can be improved. Using a large e-commerce website, for example, his product terminology may be much less than the total information data.

Thus, the present disclosure provides an example query extension method and system for realizing a normalized processing of a query.

1 and 2 illustrate an example query extension method according to a first exemplary embodiment of the present disclosure.

At 102, a query entered by a user is obtained.

At 104, a normalized query of a query is determined according to that query.

At 106, a normalized query is used as the extension term of the query that enforces the query extension.

For example, at 104, the operation may include the following.

At 1020, session information in the user's search log is obtained.

Session information refers to information describing a series of actions of a user over a successive period of time (usually between minutes and hours) on a website. During the entire process, from when the user starts browsing the website page to stopping the tour, the website server can automatically assign a session ID to the user and record the user's behavior during that time. When the user browses back to the website page after a long interval, the website server can assign a different session ID to the user and record the user's behavior. In general, within successive look-up periods, user actions usually have a specific correlation. In other words, user actions within a session are considered correlated. The queries recorded and used in the session by the user during the search and query may then also have a specific correlation. Therefore, the present disclosure can enforce normalization processing on a query based on, for example, session information.

The website server may include a specific database for storing a user's search log including session information. In order to reduce the data capacity, a search log of a user within a certain time period can be obtained. Alternatively, a search log of users within different time periods can be obtained, thereby improving the objectivity of the data.

At 1022, all queries appearing in a single session are obtained and votes are counted for each query. In a single session, according to the appearance sequence of each query, any query that appears before a particular query is counted as one vote (for that particular query).

In one session, a user may search multiple times, and the session information may include a plurality of queries. The session information may record the search sequence of the user, which is the appearance sequence of the query. The appearance sequence of the query may be determined according to each query time recorded in the session information.

An exemplary detailed process for counting votes for each query is as follows.

The queries are ordered according to the sequence of occurrences of the query. The vote for each query is the total number of queries preceding that query.

For example, five queries a, b, c, d and e are included in one session and arranged as a, b, c, d and e according to their appearance order. According to the previous definition, any query before each query is counted as a vote for each query. In other words, for query b, its vote is 1, i.e. from a to b. For query c, the vote is 2, a to c and b to c. For the query e, the vote is 4, that is, the votes from a, b, c and d respectively. In other words, the vote of each query is the total number of queries that appear before each query.

At 1024, the degree of vote similarity between a single query and the target query is determined according to the total number of votes of the target query and the number of votes of the single query for the target query in all sessions.

When multiple users visit a website within the same time period, there may be multiple sessions. The method of counting votes for each query in another session may be the same as the method described in 1022.

The total votes of the target query in all sessions can be counted as follows:

At A1, a session containing a target query is obtained.

In A2, the number of votes of the target query is obtained in each session.

In A3, the number of votes is accumulated in each session to obtain the total number of votes of the target query.

According to the above definition of a vote, any query that appears before the target query in each session provides a vote for the target query once. For example, in each session each query can be counted as one vote at most for the target query. If a query appears before the target query, the query gives a vote to the target query, otherwise the query does not give a vote to the target query. Therefore, when determining the number of votes of a single query for a target query, this may be determined by determining the number of sessions that include both the single query and the target query and where the single query appears before the target query. Exemplary detailed operations are as follows.

At B1, sessions are obtained that include both a single query and a target query.

In B2, it is determined whether a single query provides a vote for the target query in each session. If the decision is positive, then each session is selected.

At B3, the number of all selected sessions is counted to obtain the number of votes in a single query for the target query.

Obtaining a session containing a target query or a session containing both a single query and a target query can be implemented by matching. That is, a target query and / or a single query are first determined, and the determined term is matched with all queries included in the session. If the terms match, this session is determined to include the target query or two single queries and the target query.

If a specific number of votes can be determined, other methods of counting the total number of votes and the number of votes in a single query for the target query can be used.

For example, the votes of each query in each session with respect to the target query may be represented by a route scheme. The total number of votes can be calculated by counting the total route. Votes in a single query for the target query can be calculated and represented in the same route manner, and the route matches all previously counted routes. If the route matches completely, one vote is counted. The number of fully matched routes is the number of votes in a single query for the target query.

The degree of similarity between votes between a single query and the target query can be determined directly according to the ratio of the number of votes in the single query to the total number of votes in the target query. In other words, the specific ratio value is the degree of vote similarity. The higher the ratio of the total number of votes in a single query to the target query for the target query, the more users adopt route changes from the single query to the target query and the higher the degree of vote similarity between the single query and the target query.

Votes for a single query for a target query may be different in different sessions. For example, a single query can be converted directly into a target query or multiple times into a target query. For example, the query may be a, b, and c in one session, and the query may be a and c in another session. In counting, the two sessions contain votes a to c, but one of them is separated by b, while the other is not. Therefore, the two votes a to c have some difference. In order to more objectively calculate the degree of correlation between a single query and a target query, the following example operation may be implemented.

The weight and radix of each vote for the target query are determined. Each vote score is calculated according to the weight and the cardinality. The ratio of the total vote score of a single query to the target query to the total vote score of all queries for the target query is used as the degree of vote similarity between the single query and the target query.

During a particular calculation, a method of multiplying the odds of each vote by weight is adopted to calculate each vote score, and finally the scores are added to obtain the total score. For example, assume that the vote base is 1, and if a query is directly converted to a target query in one session, the weight of that query can be set as 1, and the final result of the vote remains 1 to be. If a query in another session is converted to a target query after skipping other terms, the weight of that query can be set as 0.9, then the final result of the vote is 0.9. As another example, each time a transition from a query to a target query is recorded as a step length, the weight may be determined as the inverse of the step length.

Any other method of determining the weight can be used if it can show the difference between the different votes.

At 1026, the degree of correlation between the single query and the target query is determined according to the degree of vote similarity.

In one example embodiment, the degree of vote similarity between a single query and the target query is the degree of correlation between the two.

This example embodiment only considers one dimension, the degree of vote similarity, when determining the degree of correlation. In some example embodiments, some other dimension may need to be considered when determining the degree of correlation. The values of the degree of vote similarity and the numerical values of different dimensions can be normalized to have the same quantitative level to determine a relatively accurate degree of correlation.

At 1028, the normalized query of the target query is determined according to the degree of correlation between the single query and the target query.

The threshold of the normalized query can be predetermined. That is, if the correlation degree value between the single query and the target query exceeds the threshold of the normalized query, the single query is determined as the normalized query of the target query.

In addition, after a single query is determined as a normalized query of the target query, the single query can be further refined according to a particular degree of correlation value. For example, a normalized query can be refined to synonym normalized queries, correlated normalized queries, extended normalized queries, and the like. The value range of the categories may be determined respectively. If the correlation degree value is within a specific value range, the category corresponding to the specific value range is determined as the detailed category of the normalized query. For example, if the degree of correlation between a single query and the target query is within a range of values corresponding to a synonym normalized query, the single query is determined as a synonym normalized query of the target query; If the degree of correlation between the single query and the target query is within a range of values corresponding to the correlation normalized query, the single query is determined as the correlation normalized query of the target query; If the degree of correlation between the single query and the target query is within the range of values corresponding to the synonym extended normalized query, the single query is determined as the extended normalized query of the target query.

The above-described method performs normalization processing on the query based on the information in the session. A session can record route changes or conversions of queries of different users in one query process. Objective analysis of user behavior can result in objective and accurate normalization processing.

In order to better enforce the normalization of queries, analysis from other dimensions may be performed in addition to the information in the session. Normalized processing results can be comprehensively obtained by providing preference to information obtained from sessions and other dimensions, thereby improving the objectivity of the processing results. For example, user click behavior may be analyzed or descriptive information of a query stored in the system may be analyzed.

3 illustrates another exemplary query extension method according to the second exemplary embodiment of the present disclosure. Corresponding to the operation of 104 in the first exemplary embodiment, there are the following operations before 1026 (corresponding to 308 in FIG. 3).

In 302, click information of a search result is obtained from a user's search log.

After searching using a specific query and obtaining a search result set, a user can usually click on a specific search result in the search result set. The click information of the search result may include a clicked search result, title and description information of the clicked search result, and the like.

At 304, a search result is extracted that includes the target query in the click information.

The search results may include a query that indicates the products sold in these results. Therefore, the query for obtaining each search result can be determined by analyzing the title and description information of each search result in the click information. The target query can then be extracted according to actual needs. For example, a user can search using a query called "mobile phone" and get a series of search results, while a query to get each search result is "iPhone ™", "Samsung ™ mobile phone" and "Nokia ™ mobile". Phone "and the like. If "iPhone ™" is used as the target query, all search results including "iPhone ™" can be extracted.

At 306, click similarity between the single query and the target query is determined according to the total number of clicks of all search results including the target query and the number of clicks of the search results corresponding to the single query while including the target query.

The number of clicks on a search result that corresponds to a single query while including the target query is the number of clicks on the search result that include the target query in the set of search results obtained after searching using the single query.

The total click count of all search results including the target query is the total click count of all search results including the target query while corresponding to all queries.

For example, if "iPhone ™" is the target query, "mobile phone" and "smart phone" may be used as the query for searching respectively. One set of search results is obtained after searching using the query "mobile phone", where the number of clicks of the search results including "iPhone ™" is five. The number of clicks on a search result that includes the target query "iPhone ™" corresponding to a single query "mobile phone" is five. In addition, another set of search results is obtained after a search using a query called "smartphone", where the number of clicks of the search results including "iPhone ™" is 20. The number of clicks on a search result that includes a target query "iPhone ™" corresponding to a single query "smart phone" is 20. The total number of clicks for all search results containing the target query "iPhone ™" is then 25.

The total number of clicks of all search results containing the target query and the number of clicks including the target query while corresponding to a single query can be processed as follows. The ratio of the number of clicks containing the target query to the total number of clicks of all search results containing the target query, corresponding to a single query, is calculated.

This exemplary embodiment contemplates two dimensions. Thus, under two dimensions, similarity values need to be considered when the degree of correlation between a single query and a target query is determined. Thus, the operations of 1010 in FIG. 2 corresponding to 308 of FIG. 3 need to be modified correspondingly according to determining the degree of correlation between the single query and the target query according to the degree of vote similarity and the degree of click similarity.

For example, a larger value between vote similarity and click similarity may be used as the degree of correlation between a single query and a target query.

As another example, the weight of the degree of vote similarity and the weight of the degree of click similarity are respectively determined. The degree of correlation between a single query and the target query is obtained by calculating not only the degree of vote similarity and the degree of click similarity, but also their respective weights according to one or more predetermined rules.

4 illustrates another example query extension method according to the third example embodiment of the present disclosure. For the operation of 104 in the first exemplary embodiment and the corresponding operation in the second exemplary embodiment, the operation of 1026 in FIG. 2 (corresponding to 408 of FIG. 4) is as follows. There are the same operations.

At 402, seller data stored in the server is obtained. Seller data refers to product description information specified when the seller describes the product.

At 404, seller data is analyzed and feature terms of the query as well as the query are extracted. The characteristic term of a query is a term that describes the characteristic of the query.

At 406, the degree of characteristic similarity is determined according to the characteristic term of the single query and the characteristic term of the target query.

For example, the degree of property similarity can be determined as follows. In other words,

The characteristic value of each characteristic term is determined. Feature similarity between a single query and a target query is determined by the property value.

The characteristic value of each characteristic term can be obtained by calculating mutual information between the characteristic term and the corresponding query. An exemplary calculation formula is as follows.

CP stands for query and Word stands for descriptive term. P ( CP ) and P ( Word ) represent the probability of each of the two terms appearing independently in the data set. P (CP & Word) represents the probability that two terms appear in the data set. C ( CP ) and C (Word) represent the number of pieces of information each of which the two terms appear independently in the data set. C ( CP & Word ) refers to the number of information that two terms appear together in a data set. N represents the total number of information in the data set.

Calculating the degree of characteristic similarity between a single query and a target query according to the characteristic value may be considered to calculate the degree of similarity between characteristic terms of each dimension of the single query and the target query to obtain the degree of characteristic similarity. For example, the cosine similarity degree can be used to calculate the degree of characteristic similarity. The higher the final value, the higher the degree of similarity between the two.

According to the foregoing description, since the seller data is added as a new dimension, the added dimension should be considered while determining the degree of correlation between the single query and the target query.

It can be appreciated that the new dimension can be combined with the first embodiment or with the second embodiment. That is, both dimensions and all three dimensions can be considered. When the new dimension is combined with the first embodiment, the operation of 1026 in the first embodiment corresponding to the operation of 408 may be correspondingly modified as follows. The degree of correlation between a single query and the target query is determined according to the degree of vote similarity and the degree of feature similarity.

When the three dimensions are combined, the operation of 1026 in the first embodiment corresponding to the operation of 408 may be correspondingly modified as follows. The degree of correlation between a single query and the target query is determined according to the degree of vote similarity, the degree of click similarity, and the degree of characteristic similarity.

Similar methods of the second embodiment may be referenced to determine the degree of correlation between a single query and a target query. For example, a high similarity value can be selected as the degree of correlation. Alternatively, the weight of each degree of similarity can be determined separately, and then a pre-determined method such as linear fitting can be used for the calculation based on the weight and the value. For purposes of simplicity, details are not discussed in this application.

As another example, in addition to the degree of correlation, other factors may be considered when the normalization process is performed, so that the processing result may be more accurate. For example, the degree of semantic similarity between a single query and the target query, or the degree of category similarity between the single query and the target query may be further considered.

Therefore, the following operations may be performed before operation 1028 in the first exemplary embodiment, the second exemplary embodiment, and / or the third exemplary embodiment.

The degree of semantic similarity between a single query and the target query is determined. For example, the degree of semantic similarity between a single query and a target query may be determined according to the editing distance. The edit distance is the minimum number of editing operations that convert one string to another string (or query). A Levenshtein distance is an example of an editing distance. Editing operations in Levensitine include replacing one character with another character, inserting one character, and deleting one character. The editing distance between the two strings may be calculated through a dynamic planning method. After the edit distance is calculated, a normalization process is applied to the edit distance to obtain a degree of semantic similarity so that semantic similarity can have a quantitative level equal to the degree of correlation, thereby providing convenience of subsequent processing.

Alternatively, the following operations may be performed before operation 1028 in the first exemplary embodiment, the second exemplary embodiment, and / or the third exemplary embodiment.

Category similarity between a single query and a target query is determined.

For example, in an e-commerce website, categories may be predefined to classify and manage products. Each query has a category to which it belongs. In general, categories can be divided into multiple levels. That is, one first level category may include a plurality of second level categories, and each second level category may be further divided into a plurality of third level categories and the like. The degree of category similarity may be determined by determining whether the queries belong to the same first level category, second level category, third level category, or the like. Similarly, normalization processing is applied to the degree of category similarity during the decision process such that the degree of category similarity has the same quantitative level as the degree of correlation.

As one or more new factors are considered, the operation of 1028 in the three embodiments described above may be correspondingly modified.

If the degree of correlation and the degree of semantic similarity are considered, the operation of 1028 may be modified to determine the normalized query of the target query according to the degree of correlation and the degree of semantic similarity between the single query and the target query.

If degree of correlation and category similarity are considered, the operation of 1028 may be modified to determine the normalized query of the target query according to the degree of correlation and category similarity between the single query and the target query.

If all three are considered simultaneously, the operation of 1028 may be modified to determine the normalized query of the target query according to the degree of correlation, degree of semantic similarity, and category similarity between the single query and the target query.

For example, during the refinement process, linear fit can be applied to two or three factors to obtain the normalization score of a single query and the target query, and the normalized query of the target query is determined according to the normalization score.

For example, with reference to the first exemplary embodiment, thresholds can be used for enforcement. In other words, the threshold of a normalized query is predetermined. If the normalization score exceeds the threshold, a single query is determined as the normalized query of the target query. Moreover, normalized queries can be further classified. In other words, different values corresponding to different categories are set respectively. When a normalized integer is in a specific value range, the category corresponding to that value range is determined as the detailed category of the normalized query. Reference may be made to the specific method in the first embodiment. For purposes of simplicity, details are not discussed in this application.

With respect to the above-described operations added before any particular operation of any exemplary embodiment, the operations so added are independent of any particular operation, and so the added operations may be performed before or before the specific operation. It may be appreciated that the operation may be performed simultaneously with or prior to such an operation. This disclosure does not impose any limitation on this application. For example, for the operations 302 to 306 added in the second exemplary embodiment, these operations may be performed between the operation of 1024 and the operation of 1028. Alternatively, these operations may be performed concurrently with the operations 1020 through 1024. Alternatively, these operations may be performed prior to the step of 1020. This disclosure does not impose any limitation on this application. Other example embodiments may have similar processing, which is not described in detail herein.

5 illustrates a first example query extension system 500 in accordance with the present disclosure. System 500 may include one or more processor (s) 502 and memory 504. Memory 504 is an example of a computer-readable medium. As used in this application, “computer-readable media” includes computer storage media and communication media.

Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer-executable instructions, data structures, program modules, or other data. In contrast, a communication medium may implement computer-readable instructions, data structures, program modules, or other data into modulated data signals such as carrier waves. As defined herein, computer storage media does not include communication media. The memory 204 may store a program unit or module and program data therein.

In the example of FIG. 5, memory 504 may store query input module 506, normalized query determination module 508, and query extension module 510 therein. The query input module 506 obtains a query entered by the user. The normalized query determination module 508 determines the normalized query of the query according to the query. The query expansion module 510 enforces query expansion using the normalized query as an extension term of the query.

6 illustrates an example normalized query determination module 600. Exemplary normalized query determination module 600 includes session information acquisition module 602, query vote calculation module 604, vote similarity determination module 606, correlation degree determination module 608, and normalized query determination module. 610 may be included.

The session information obtaining module 602 obtains session information from the search log of the user.

Query vote calculation module 604 obtains all queries appearing in a single session and counts votes for each query. In a single session, depending on the appearance sequence of each query, any query that appears before a particular query is counted as one vote (for that particular query).

The vote similarity determining module 606 determines the degree of vote similarity between the single query and the target query according to the total number of votes of the target query and the number of votes of the single query for the target query in all sessions. For example, the vote similarity degree determination module 606 may include an odd and weight determination unit, a score calculation unit, and a ratio calculation unit. The cardinal and weight determination unit determines the weight and cardinality of each vote for the target query. The score calculation unit calculates the score of each vote according to the weight and the cardinality. The ratio calculation unit uses the ratio of the total vote score of a single query to the target query to the total vote score of all the queries for the target query as the degree of vote similarity between the single query and the target query.

The degree of correlation determination module 608 determines the degree of correlation between the single query and the target query according to the degree of vote similarity.

The normalized query determination module 610 determines the normalized query of the target query according to the degree of correlation between the single query and the target query.

For example, the normalized query determination module 610 sets a threshold of the normalized query and if the correlation degree value between the single query and the target query exceeds the threshold of the normalized query, the normalized query of the target query. It may include a normalized query threshold setting unit to determine as.

As another example, the normalized query determination module 610 may also include a normalized query category classification unit, a value range setting unit, and a category determination unit. The normalized query category classification unit separates the normalized query categories into synonym normalized queries, correlated normalized queries, and extended normalized queries. The value range setting unit sets the value range of the three categories in descending order according to the correlation degree value. The category determining unit determines the category corresponding to the range of values to which the degree of correlation between the single query and the target query belongs as a detailed category of the single query and the target query.

7 illustrates a second example query extension system that includes a second example normalized query determination module 700. In addition to the session information acquisition module 602, the query vote calculation module 604, the vote similarity determination module 606, the correlation degree determination module 608, and the normalized query determination module 610, normalized query determination. The module 700 further includes a click information acquisition module 702, a search result extraction module 704, and a click similarity degree determination module 706. The click information obtaining module 702 obtains click information of a search result from a user's search log. The search result extraction module 704 extracts a search result including the target query from the click information. The click similarity determination module 706 determines the degree of click similarity between a single query and the target query according to the total number of clicks of the search results including the target query and the total number of clicks of the search results corresponding to a single query while including the target query. do. The degree of correlation determination module 610 determines a degree of correlation between the single query and the target query according to the degree of vote similarity and the degree of click similarity.

8 illustrates a third example query extension system that includes a third example normalized query determination module 800. In addition to the session information acquisition module 602, the query vote calculation module 604, the vote similarity determination module 606, the correlation degree determination module 608, and the normalized query determination module 610, normalized query determination. Module 800 further includes merchant data acquisition module 802, data analysis module 804, and feature similarity degree determination module 806.

The seller data acquisition module 802 acquires seller data stored in the server. Seller data includes product description information specified when the seller describes the product.

The data analysis module 804 analyzes seller data and extracts query terms as well as feature terms from the seller data.

The degree of characteristic similarity determining module 806 determines characteristic similarity according to characteristic terms of a single query and a target query. For example, the characteristic similarity degree determining module 806 may include a characteristic value calculation unit that calculates the characteristic value of each characteristic term. The characteristic value is calculated according to the mutual information of the characteristic term and its corresponding query.

The degree of correlation determination module 608 determines the degree of correlation between the single query and the target query according to the degree of vote similarity and the degree of characteristic similarity.

It can be appreciated that the relevant data described in the second and third exemplary embodiments can be further considered simultaneously for processing. That is, the correlation degree determining module 608 may further determine the degree of correlation between the single query and the target query according to the degree of vote similarity, the degree of click similarity, and the degree of characteristic similarity. While determining the degree of correlation in detail, the largest of the three may be used as the degree of correlation. Alternatively, linear fit can be applied to three degrees of similarity to obtain the final value as the degree of correlation.

When the degree of vote similarity is combined with the degree of click similarity or the degree of characteristic similarity, it is convincing that the larger of the two can be selected as the degree of correlation. Alternatively, linear fit can be applied to two degrees of similarity to find the final value as the degree of correlation.

As another example, one or more of the example systems may further include a semantic similarity degree determining module and / or a category similarity degree determining module that respectively determine a degree of semantic similarity and / or category similarity between a single query and a target query.

Correspondingly, the normalized query determination module 610 may determine the normalized query of the target query according to the degree of correlation and semantic similarity between the single query and the target query, or the degree of correlation and category similarity between the single query and the target query. The normalized query of the target query may be determined or the normalized query of the target query may be determined according to the degree of correlation, the degree of semantic similarity, and the category similarity between the single query and the target query.

For example, the semantic similarity degree determination module may include an edit distance determination unit and a normalization processing unit. The edit distance determining unit determines the edit distance between a single query and the target query. The editing distance refers to the minimum number of editing operations for converting one term into another term. The normalization processing unit performs the normalization processing of the editing distance to obtain a degree of semantic similarity having the same quantitative level as the correlation degree.

All example embodiments are described in the present disclosure. Each exemplary embodiment emphasizes differentness from the other example embodiments. The same or similar parts of the example embodiments may be cross referenced. Because the example system embodiments are basically similar to the example method embodiments, the example system embodiments are not described in detail. Related parts may be referred to as part of exemplary embodiments.

Some example query extension methods and systems of the present disclosure are not described in detail herein. The present disclosure describes the principles and implementation of the present technology using some examples. Example embodiments are merely intended to assist in understanding the methods and key concepts of the present disclosure. On the other hand, those of ordinary skill in the art may modify or change the exemplary embodiments or applications in accordance with the concepts of the present disclosure, which will still fall within the protection scope of the present disclosure. This disclosure should not be understood as a limitation on the present technology.

Claims (20)

Obtaining a target query,
Determining a normalized query according to the obtained target query,
The determining step,
Obtaining session information from a search log;
Determining a vote similarity degree between a single query and the target query based on the session information, wherein determining the vote similarity degree comprises:
Obtaining all queries appearing in a single session, including the target query;
Calculating a vote count for each target query, wherein the calculating includes counting the single query appearing before the target query in the single session as a vote from the single query for the target query; and ,
Determining a degree of vote similarity between the single query and the target query according to the calculated number of votes;
Determining a correlation degree between the single query and the target query based in part on the degree of vote similarity;
Determining the normalized query based in part on the degree of correlation between the single query and the target query.
Way.
The method of claim 1,
Using the normalized query as an extended term of the obtained target query;
Way.
The method of claim 1,
The determining the degree of similarity between votes between the single query and the target query according to the calculated number of votes is based on the total number of votes of the target query and the number of votes of the single query for the target query in every session.
Way.
The method of claim 3, wherein
Computing the total number of votes of the target query in all sessions, wherein calculating the total number of votes of the target query,
Acquiring one or more sessions comprising the target query;
Counting the number of votes of the target query in each session;
Accumulating the number of votes of the target query in each session to obtain a total number of votes of the target query;
Way.
The method of claim 3, wherein
Computing the number of votes of the single query for the target query, wherein calculating the number of votes of the single query for the target query,
Acquiring one or more sessions including the single query and the target query;
Determining whether the single query provides a vote for the target query in each session;
In response to determining that the single query provides a vote for the target query in each session, selecting each session;
Counting the number of selected sessions to obtain a vote count of the single query for the target query;
Way.
The method of claim 1,
Determining the degree of similarity between votes between the single query and the target query according to the calculated number of votes may be performed by comparing the ratio of the total number of votes of the single query to the total number of votes of the target query to the target query. Using as a degree of vote similarity between the target queries;
Way.
The method of claim 1,
Determining the degree of vote similarity between the single query and the target query,
Determining a weight and a base number of each vote of the target query;
Calculating a score for each vote according to each of the weights and each of the cardinalities,
Using the ratio of the total vote scores of the single query to the target query to the total vote scores of all the queries for the target query as a degree of vote similarity between the single query and the target query.
Way.
The method of claim 1,
Determining the normalized query based in part on the degree of correlation between the single query and the target query,
Setting thresholds for normalized queries,
In response to determining that a value of the degree of correlation between the single query and the target query exceeds a threshold of the normalized query, determining the single query as a normalized query of the target query.
Way.
The method of claim 1,
Determining the normalized query based in part on the degree of correlation between the single query and the target query,
Separating the categories of normalized queries into synonym normalized queries, correlated normalized queries, and extended normalized queries;
Setting a range of values of each of the synonym normalized query, the correlated normalized query, and the extended normalized query in descending order according to a correlation degree value;
Using a category corresponding to a range of values to which a degree of correlation between the single query and the target query belongs as a detailed category of the single query and the target query;
Way.
The method of claim 1,
Acquiring click information of a search result from the search log;
Extracting one or more search results including a target query according to the click information;
Determining a degree of click similarity between the single query and the target query according to the total number of clicks of the search result including the target query and the number of clicks of the search result including the target query while corresponding to the single query. and,
Determining a degree of correlation between the single query and the target query is based on the degree of vote similarity and the degree of click similarity.
Way.
The method of claim 10,
Determining a degree of correlation between the single query and the target query based on the degree of vote similarity and the degree of click similarity is the largest value between the degree of vote similarity and the degree of click similarity between the single query and the target query. Selecting as the degree of correlation
Way.
The method of claim 10,
Determining a degree of correlation between the single query and the target query based on the degree of vote similarity and the degree of click similarity,
Determining weights of the vote similarity degree and the click similarity degree, respectively;
Calculating a degree of correlation between the single query and the target query according to one or more predetermined rules according to the degree of vote similarity and the degree of click similarity and their respective weights.
Way.
The method of claim 1,
Obtaining seller data stored in a server, wherein the seller data includes product description information; and
Analyzing the seller data to extract a query and characteristic terms of the query;
Determining a degree of characteristic similarity according to each characteristic term of the single query and the target query,
Determining the degree of correlation between the single query and the target query is based on the degree of vote similarity and the degree of feature similarity.
Way.
The method of claim 13,
Determining the degree of characteristic similarity according to each characteristic term of the single query and the target query,
Calculating characteristic values of each characteristic term based on mutual information between each characteristic term and a query corresponding to each characteristic term;
Calculating a degree of characteristic similarity between the single query and the target query according to the characteristic value.
Way.
The method of claim 1,
Determining a degree of semantic similarity between the single query and the target query,
Determining the normalized query of the target query is based on the degree of correlation and the degree of semantic similarity between the single query and the target query.
Way.
The method of claim 1,
Determining a degree of category similarity between the single query and the target query,
Determining the normalized query of the target query is based on the degree of correlation and category similarity between the single query and the target query
Way.
The method of claim 1,
Determining a degree of semantic similarity between the single query and the target query;
Determining a degree of category similarity between the single query and the target query,
Determining the normalized query of the target query is based on the degree of correlation, the degree of semantic similarity, and the category similarity between the single query and the target query.
Way.
Acquiring session information from a search log;
Determining a degree of vote similarity between a single query and a target query based on the session information, wherein determining the degree of vote similarity
Obtaining all queries appearing in a single session, including the target query;
Calculating a vote count for each target query, wherein the calculating includes counting the single query appearing before the target query in the single session as a vote from the single query for the target query; and ,
Determining the degree of vote similarity between the single query and the target query according to the calculated number of votes;
Determining a degree of correlation between the single query and the target query based on the degree of vote similarity;
Determining a degree of semantic similarity between the single query and the target query;
Determining a normalized query based on the degree of correlation and the degree of semantic similarity between the single query and the target query.
Way.
The method of claim 18,
Determining the degree of semantic similarity between the single query and the target query,
Determining an editing distance between the single query and the target query, wherein the editing distance is the minimum number of editing operations changing from one term to another term; and
Normalizing the editing distance to obtain the degree of semantic similarity having a quantitative level equal to the degree of correlation;
Way.
As a system for extending queries,
One or more processors,
A memory device communicatively coupled with the one or more processors;
A query input module having instructions executable by the one or more processors to obtain a target query;
A normalized query determination module having instructions executable by the one or more processors to determine a normalized query in accordance with the obtained target query;
A query extension module having instructions executable by the one or more processors to implement query extension using the normalized query as an extension term of the obtained target query;
Determining the normalized query according to the obtained target query,
Obtaining session information from the search log,
Determining a degree of vote similarity between a single query and a target query based on the session information-determining the degree of vote similarity,
Obtaining all queries appearing in a single session, including the target query;
Calculating a number of votes for each target query, the calculating comprising counting the single query appearing before the target query in the single session as a vote from the single query for the target query; and
Determining the degree of vote similarity between the single query and the target query according to the calculated number of votes; and
Determining a degree of correlation between the single query and the target query,
Determining a normalized query based in part on the degree of correlation between the single query and the target query.
system.

Images